Your Code Must Be Faster

tl;dr: As agents take over writing code, the cost of validating generated code becomes a primary determinant of engineering throughput. Optimizing your test loop to minimize RAM, CPU, and wall clock time now yields multiplicative returns.

Writing code is an iterative hypothesis testing loop: write some code you think works, run it, see how it fails.

Agents have not escaped this process: like human developers, their velocity is constrained it.

If an agent takes five minutes to write code and 55 minutes to test it, it can test one hypothesis an hour. If it takes 5 minutes to write code and 5 minutes to test it, it can test 6 hypotheses an hour.

What’s changed is that you can run agents in parallel. 10 agents at one hypothesis an hour - ten hypotheses an hour. 10 agents at 6 hypotheses an hour - 60 hypotheses an hour.

Thus: compile time and code & test execution time are paramount for delivery in a world with agents.

There’s another variable, though: resource use. More resource use constrains how many tests you can run at once. If you have 20GB of RAM, and your application takes 15GB, you can’t parallelize. Ditto for resource utilization.

Optimal technology choices are now weighted towards efficiency more than ever before.